diff options
Diffstat (limited to 'src/ollama_web/llm.py')
| -rw-r--r-- | src/ollama_web/llm.py | 55 |
1 files changed, 38 insertions, 17 deletions
diff --git a/src/ollama_web/llm.py b/src/ollama_web/llm.py index 76a61a5..2839051 100644 --- a/src/ollama_web/llm.py +++ b/src/ollama_web/llm.py @@ -30,6 +30,10 @@ MAX_TOOL_CONTEXT_CHARS = 24000 # Type alias matching ollama's expected tools signature. ToolCallable = Callable[..., Any] +# Marker that separates a paged tool result's body from its page-info footer. +# Must match ``ollama_web.tools.helper.pagination.PAGE_INFO_MARKER``. +_PAGE_INFO_MARKER = "\n---page-info---" + def _safe_tool_arguments(arguments: Any) -> Any: return redact_secrets(arguments) @@ -40,8 +44,6 @@ def _safe_tool_result(result: str) -> str: return str(redacted) - - def get_client(host: str | None = None) -> ollama.Client: """Return an ``ollama.Client`` configured for the configured host.""" import httpx @@ -340,17 +342,40 @@ def chat_with_tools( yield {"type": "thinking", "content": thinking} if content: yield {"type": "delta", "content": content} - done_event = {"type": "done"} + done_event_final: dict[str, Any] = {"type": "done"} if tokens: - done_event.update(tokens) - yield done_event + done_event_final.update(tokens) + yield done_event_final def _trim_tool_result(result: str, limit: int) -> str: - """Truncate ``result`` to ``limit`` chars while keeping it readable.""" + """Truncate ``result`` to ``limit`` chars while keeping it readable. + + When the result contains a ``---page-info---`` footer (emitted by paged + tools), the footer is preserved and only the body is truncated so the + model can still see the ``next_token`` needed to fetch the continuation. + """ if len(result) <= limit: return result - # Prefer cutting at a paragraph boundary. + + # Protect the page-info footer from truncation. ``split`` keeps the + # marker on the footer side so the ``---page-info---`` line is preserved. + if _PAGE_INFO_MARKER in result: + body, footer = result.split(_PAGE_INFO_MARKER, 1) + footer = _PAGE_INFO_MARKER + footer + body_limit = limit - len(footer) + if body_limit <= 0: + # The footer alone exceeds the limit; keep it verbatim and drop the + # body entirely so the model can still continue paging. + return "…[tool result truncated]" + footer + if len(body) <= body_limit: + return result + cutoff = body[:body_limit].rsplit("\n\n", 1)[0] + if len(cutoff) < body_limit * 0.5: + cutoff = body[:body_limit].rsplit(" ", 1)[0] + return cutoff + "\n\n…[tool result truncated]" + footer + + # No footer: legacy truncation behaviour. cutoff = result[:limit].rsplit("\n\n", 1)[0] if len(cutoff) < limit * 0.5: # No good paragraph boundary; cut at a space instead. @@ -457,9 +482,7 @@ def stream_chat_with_tools( # stream with almost no meaningful content. Log it and retry a few # times before giving up. is_empty = ( - not full_message.get("content") - and not full_message.get("thinking") - and not tool_calls + not full_message.get("content") and not full_message.get("thinking") and not tool_calls ) if is_empty and empty_retries < MAX_EMPTY_RETRIES: empty_retries += 1 @@ -497,7 +520,7 @@ def stream_chat_with_tools( "type": "warning", "message": t("status.empty_response", lang=language), } - done_event = {"type": "done"} + done_event: dict[str, Any] = {"type": "done"} tokens = _token_counts(last_chunk) if tokens: done_event.update(tokens) @@ -594,11 +617,11 @@ def stream_chat_with_tools( if content: yield {"type": "delta", "content": content} logger.info("final answer stream finished elapsed=%.2fs", time.time() - t0) - done_event = {"type": "done"} + done_event_final: dict[str, Any] = {"type": "done"} tokens = _token_counts(last_chunk) if tokens: - done_event.update(tokens) - yield done_event + done_event_final.update(tokens) + yield done_event_final async def astream_chat_with_tools( @@ -690,9 +713,7 @@ def _assistant_message_chunk(full_message: dict[str, Any]) -> Message | None: name = c.get("name", "") args = c.get("arguments", {}) tc_list.append( - Message.ToolCall( - function=Message.ToolCall.Function(name=name, arguments=args) - ) + Message.ToolCall(function=Message.ToolCall.Function(name=name, arguments=args)) ) kwargs["tool_calls"] = tc_list return Message(**kwargs)
\ No newline at end of file |